Co-Scheduling HPC Workloads on Cache-Partitioned CMP Platforms

Aupy, Guillaume; Benoît, Anne; Goglin, Brice; Pottier, Loïc; Robert, Yves

doi:10.1109/cluster.2018.00052

Cited by 10 publications

(6 citation statements)

References 25 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…J. Breitbart et al created a resource monitoring tool useful for co-scheduling HPC applications [16] and provided a memory-intensity-aware co-scheduling policy [17]. Since the industry started supporting several QoS control features, some researchers combined the above concepts with cache partitioning [18], [19], bandwidth partitioning [20], [21] or the combination of them [22], [23]. Q. Zhu et al rather targeted CPU-GPU heterogeneous processors and proposed a co-scheduling approach suitable for them [24].…”

Section: Related Workmentioning

confidence: 99%

Hierarchical Resource Partitioning on Modern GPUs: A Reinforcement Learning Approach

Saroliya,

Arima,

Liu

et al. 2023

2023 IEEE International Conference on Cluster Computing (CLUSTER)

View full text Add to dashboard Cite

GPU-based heterogeneous architectures are now commonly used in HPC clusters. Due to their architectural simplicity specialized for data-level parallelism, GPUs can offer much higher computational throughput and memory bandwidth than CPUs in the same generation do. However, as the available resources in GPUs have increased exponentially over the past decades, it has become increasingly difficult for a single program to fully utilize them. As a consequence, the industry has started supporting several resource partitioning features in order to improve the resource utilization by co-scheduling multiple programs on the same GPU die at the same time.Driven by the technological trend, this paper focuses on hierarchical resource partitioning on modern GPUs, and as an example, we utilize a combination of two different features available on recent NVIDIA GPUs in a hierarchical manner: MPS (Multi-Process Service), a finer-grained logical partitioning; and MIG (Multi-Instance GPU), a coarse-grained physical partitioning. We propose a method for comprehensively co-optimizing the setup of hierarchical partitioning and the selection of co-scheduling groups from a given set of jobs, based on reinforcement learning using their profiles. Our thorough experimental results demonstrate that our approach can successfully set up job concurrency, partitioning, and co-scheduling group selections simultaneously. This results in a maximum throughput improvement by a factor of 1.87 compared to the time-sharing scheduling.

show abstract

Section: Related Workmentioning

confidence: 99%

Hierarchical Resource Partitioning on Modern GPUs: A Reinforcement Learning Approach

Saroliya,

Arima,

Liu

et al. 2023

2023 IEEE International Conference on Cluster Computing (CLUSTER)

View full text Add to dashboard Cite

show abstract

“…The latter seeks to partition a graph into even-sized components while minimizing the number of edges. Figure 1 summarizes other problems that can be modeled as Balanced Partition, including: People Assignment [34], Routing [31], [35]- [38], Task Allocation [39], [40], File Placement [41], [42] and Scheduling [43]- [45]. We briefly describe them in the following lines.…”

Section: A Balanced Partition (Bp)mentioning

confidence: 99%

“…Here, the goal is to improve the parallel execution efficiency of tasks that share resources (e.g., external devices, shared memory, and files). To solve this problem, there exist multiple paradigms, such as a Partitioned Schedule [43], [45], a Global Schedule [46], and a Semi-partitioned Schedule [44].…”

Section: A Balanced Partition (Bp)mentioning

confidence: 99%

Sequence-Based Selection Hyper-Heuristic Model via MAP-Elites

et al. 2021

View full text Add to dashboard Cite

Combinatorial optimization problems (COPs) are paramount for real-life problems with discrete variables. Even though the number of combinations is finite, some problems exhibit exponential growth, rendering exact approaches unfeasible. So, approximate methods, such as heuristics, are customary for making fast decisions. Despite their small computational cost, each heuristic specializes in specific kinds of problems. Hence, other approaches have appeared for merging their strengths. One of these approaches is commonly known as selection hyper-heuristics. However, a trained hyper-heuristic usually carries the following problem: it provides scarce information about its sensitivity. This makes it more difficult to estimate the effect of changing a parameter. Illumination algorithms seek to improve this issue by focusing on exploration rather than on exploitation while preserving information about the best solutions with different criteria. Still, literature falls short when merging both approaches, representing a knowledge gap. Therefore, in this work, we test the feasibility of using an illumination algorithm, the MAP-Elites (ME), for tuning a sequence-based selection hyper-heuristic model for Balanced Partition problems. We choose ME since it has been successfully applied to a different COP, and by linking it with hyper-heuristics, we may exploit upon its benefits. So, we may achieve a hyper-heuristic that represents the best combination of heuristics. Simultaneously, we can gain information about the expected performance of diverse hyperheuristics, e.g., those that begin with a different heuristic. Our approach operates by creating a multidimensional map, where each design variable represents the application of a given heuristic. Afterward, ME generates mutated sequences and tests them to determine if they represent a better-performing solution. To test our approach, we consider 1500 instances that include easy and hard instances, analyzed under different scenarios. We also include limit instances that are neither easy nor hard. Our resulting data supports the feasibility of the proposed approach. We show that this kind of hyper-heuristic can perform toe-to-toe with a synthetic oracle. Besides, under some conditions, the former may even outperform the latter. This represents an outstanding result, since such an oracle can only be obtained by a brute-force approach, and because it demonstrates that merging both approaches (ME and hyper-heuristics) is a path worth pursuing. We also present how each parameter affects the model performance, concluding that some parameters can be critical while others are virtually irrelevant. This is also a relevant result, since it serves as the groundwork for future works so that they can focus on exploiting the most relevant parameters.

show abstract

“…Some authors in their publications approach it as an offline scheduling problem, where all task data are available at the start and an optimal schedule can be constructed in advance. Among these publications are [14,20,21], where the authors solve the offline scheduling problem with resource constraint. They model the CPU cache partition size as a controllable task resource and define task speedup as a function of the cache partition size.…”

Section: Related Workmentioning

confidence: 99%

Analytical and Numerical Evaluation of Co-Scheduling Strategies and Their Application

Kuchumov

Korkhov

2021

Computers

View full text Add to dashboard Cite

Applications in high-performance computing (HPC) may not use all available computational resources, leaving some of them underutilized. By co-scheduling, i.e., running more than one application on the same computational node, it is possible to improve resource utilization and overall throughput. Some applications may have conflicting requirements on resources and co-scheduling may cause performance degradation, so it is important to take it into account in scheduling decisions. In this paper, we formalize the co-scheduling problem and propose multiple scheduling strategies to solve it: an optimal strategy, an online strategy and heuristic strategies. These strategies vary in terms of the optimality of the solution they produce and a priori information about the system they require. We show theoretically that the online strategy provides schedules with a competitive ratio that has a constant upper limit. This allows us to solve the co-scheduling problem using heuristic strategies that approximate this online strategy. Numerical simulations show how heuristic strategies compare to the optimal strategy for different input systems. We propose a method for measuring input parameters of the model in practice and evaluate this method on HPC benchmark applications. We show the high accuracy of the measurement method, which allows us to apply the proposed scheduling strategies in the scheduler implementation.

show abstract

Co-Scheduling HPC Workloads on Cache-Partitioned CMP Platforms

Cited by 10 publications

References 25 publications

Hierarchical Resource Partitioning on Modern GPUs: A Reinforcement Learning Approach

Hierarchical Resource Partitioning on Modern GPUs: A Reinforcement Learning Approach

Sequence-Based Selection Hyper-Heuristic Model via MAP-Elites

Analytical and Numerical Evaluation of Co-Scheduling Strategies and Their Application

Contact Info

Product

Resources

About